Over the last decade, AI and machine learning have become increasingly valuable in the geophysical sciences. However, AI is often seen as untrustworthy due to its "black-box" nature because the decision-making process is not easily understood. This lack of transparency in AI-generated results creates uncertainties and mistrust in the science, which can dampen the impact of results. Explainable AI addresses the transparency issue by providing a good understanding of deep-learning models and their results.
More recently, creative uses of xAI have been proven helpful in attributing anthropogenic climate change to extreme precipitation and temperature events. xAI methods like counterfactuals and anchors allow researchers to understand deep-learning model decision-making boundaries. Other methods, such as backwards optimization and salcieny maps provide attribution of features in the data that are important to a particular decision. Although xAI methods are relatively new to the geosciences, they have proved extremely useful in boosting scientific discovery and helping people better understand deep learning.
Relevant Research
- Using Explainable Artificial Intelligence to Quantify “Climate Distinguishability” After Stratospheric Aerosol Injection
- Trustworthy Artificial Intelligence for Environmental Sciences: An Innovative Approach for Summer School
- Carefully choose the baseline: Lessons learned from applying XAI attribution methods for regression tasks in geoscience
- Investigating the fidelity of explainable artificial intelligence methods for applications of convolutional neural networks in geoscience
Climate downscaling using artificial intelligence (AI) represents a cutting-edge approach to enhance the resolution and accuracy of climate projections derived from global climate models (GCMs). Traditional dynamical downscaling techniques are computationally expensive, and statistical methods often fail to capture the complexity of local climate variations influenced by topography, land use, and atmospheric processes. AI-based models, such as deep learning networks and generative models, offer a powerful alternative by learning intricate relationships between coarse-scale GCM outputs and fine-scale regional climate data. Techniques like Convolutional Neural Networks (CNNs), Generative Adversarial Networks (GANs), and Diffusion Models have shown significant promise in bridging the spatial resolution gap by generating high-resolution climate fields from low-resolution inputs, while capturing non-linear and stochastic processes inherent in climate systems. AI-driven climate downscaling models not only improve spatial resolution but also incorporate probabilistic predictions, enabling uncertainty quantification and ensemble generation, which is crucial for robust climate risk assessments. This paradigm shift toward AI-powered downscaling offers transformative potential in climate modeling, especially in areas such as extreme weather event prediction, regional climate change assessments, and climate impact modeling. The outcome of this research project will provide highly relevant and actionable downscaled climate data at local scales, empowering societal decision-making, particularly in areas of local governance and community preparedness, where accurate, high-resolution climate information is crucial for managing risks and adapting to climate change.
Downscaling of different types of storms using AI
